A Recency-Weighted Monte Carlo Engine for Possession-Level Basketball Forecasting

Abstract

This paper presents a modular Monte Carlo (MC) engine for retrospective forecasting experiments on historical National Basketball Association games. Instead of fitting a single discriminative winner classifier, the system estimates player event rates from one-year individual box-score histories and simulates games as sequences of possession-level events. The study states the simulation assumptions formally and evaluates the engine with a season-wide benchmark. The framework is mechanistic and modular: a rolling box-score estimator, position-aware role sampling, explicit possession transitions, and separable calibration produce full score distributions and permit controlled what-if experiments. On a matched set of 1106 games from the 1991–1992 regular season, the MC estimate averaged over three runs reaches an area under the receiver operating characteristic curve (AUC) of 0.677 after tie normalization. Averaging per-game probabilities across three independent 1000-epoch exports and applying a training-season home-prior logit calibration yields 0.666 accuracy, Brier score 0.216, log loss 0.623, and AUC 0.677. Training-tuned Elo and regularized logistic regression provide stronger binary forecasting metrics, whereas the MC engine provides complementary process-level transparency and distributional outputs, including score distributions, modal scores, tie rates, plausible score ranges, and uncertainty summaries.